Cognitive-Uncertainty Guided Knowledge Distillation for Accurate Classification of Student Misconceptions

Researchers proposed a two-stage knowledge distillation method to better classify student misconceptions with small models.

The system first filters high-value training samples using teacher uncertainty and confidence gaps, then adjusts hard/soft label mixing by sample difficulty.

On benchmark tests, it improves accuracy and MAP@3 while using far fewer filtered samples and only a 4B model.

The work is notable because it handles scarce, noisy educational data well and shows compact models can beat much larger ones in assessment tasks.